- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper studies contextual bandits where a free surrogate is observed after action selection and before deciding whether to acquire a costly primary outcome. ASB combines policy-comparison relevance with estimated surrogate-conditioned residual RMS, augmented inverse-propensity feedback, and a hard acquisition budget. The paper characterizes optimal PRE/POST variance under a proportional-allocation feasibility condition, separates the two information orders in regret, and proves a worst-case budget-dependent regret guarantee. Synthetic mechanism studies and a retrospective KuaiRec benchmark evaluate the design.
Critical strengths
- A meaningful information-order question with a principled acquisition rule. Policy relevance times surrogate-conditioned residual uncertainty is intuitive. Theorem 1 characterizes PRE/POST variance and strict improvement under its stated feasibility assumptions; Theorem 2 shows a bounded-versus-
Omega(T/B)regret separation even when the surrogate's marginal law is identical across possible optimal actions. This identifies value in post-action information beyond predicting which action is better. - Careful learning and acquisition under a hard budget. The support floor, interval accounting and stopped-process argument address the difficulty that budget exhaustion can invalidate unbiased feedback. The best-in-class guarantee accommodates misspecified models, and its worst-case budget dependence has a lower bound. These are useful contributions even though stronger benefits from learning accurate residual models remain unquantified.
Critical weaknesses
- The benchmarks provide limited evidence of the need to learn from the acquired outcomes. On all 64×512 released synthetic world/context pairs, the recommendation-only plurality rule
argmax_a sum_pi 1{pi(x)=a}chooses the optimal action with zero errors or ties (minimum winning margin 68). Its mapping is represented in the policy class and it needs no outcomes or surrogates. This does not invalidate the matched acquisition ablations: the full component rerun gives ASB 550.70±4.80 SE versus relevance-only 650.25±5.57. It does limit what they show about an unresolved decision problem. On KuaiRec, plurality gives cumulative regret 49.43 versus ASB means 327.59–696.83 across budgets. This is a different action learner using the same 7,055 offline fitting labels, not an acquisition-only comparison or a claim of no prior supervision. It also does not dominate terminal value: at the 24% budget ASB reaches .365476 versus plurality's .364400. The operational distinction between a free play-duration threshold and a costly watch ratio needs justification. - The benefit of the learned surrogate allocation is less fully established than the information-order separation. Residual-RMS allocation and augmentation have close precedents in the cited Active Statistical Inference and Ao et al. work; contextual regret and hard-budget integration remain distinct. Proposition 10 controls model-error effects at a fixed learner state but does not give an error-versus-acquisition-budget rate or a regret comparison between separately evolving learners. Theorem 3 likewise has no explicit surrogate-quality improvement. Its ASB exploration floor does not attain the separate constant-regret POST construction, and the experimental adaptive learning rates/restarts are not the analyzed fixed-eta instantiation. These limit the guarantees' scope; they do not contradict the stated theorems.
Questions for authors
- Benchmark and policy representation: Can you include the recommendation-only baseline, distinguish cumulative regret from terminal value, and test the acquisition benefit with a policy class whose consensus does not reveal the optimum? Since uniform expert initialization changes the induced action distribution when one behavioral cluster is overrepresented, how sensitive are the conclusions to distinct but nearly identical policies? Exact duplicates can be deduplicated; the issue is robustness to the prior induced by the policy collection, not a theorem contradiction.
- Learned models and analyzed algorithm: Which bridge/residual fitting choices produce the gains, and what evidence connects their acquisition cost and accuracy to policy regret? Can you clarify the guarantee for the adaptive-rate/restart implementation, or compare it with the analyzed schedule?
- Practical budget choices: What makes watch ratio costly when the play-duration threshold is free? Please account for the 7,055 offline labels and 346-label initialization. How sensitive are the results to the three post-initialization intervals, their lengths and the residual-uncertainty interval allocation? Reporting final-fill acquisition frequency would help interpret how much the adaptive score determines the final budget use.
Follow-up on the revised manuscript
Thank you again for your detailed comments. Following our two earlier responses, we have uploaded the revised manuscript with updates to
If you have a chance to take another look, we would appreciate your thoughts.